I work as a research group lead at the leading artificial intelligence company DeepMind and hold a part-time position as Chair of Machine Learning at the Computer Science Department of University College London (UCL). In support of responsible innovation in artificial intelligence, I also serves as a Member of the Board of Directors at the Partnership on AI (PAI).
I studied physics at the University of Hamburg, at Imperial College London, and at the Technical University of Berlin, where I also obtained my PhD in machine learning in 2001. After holding post-doctoral positions at ETH Zurich and Royal Holloway College, University of London, I joined Microsoft Research in Cambridge in 2003, where I co-founded the Online Services and Advertising group. Major applications of my work include Xbox Live’s TrueSkill system for ranking and matchmaking and the AdPredictor framework for click-through rate prediction in Bing. Furthermore, my work on the predictability of private attributes from digital records of human behaviour has been the subject of intense discussion among privacy experts and the general public. At DeepMind, I have returned to my original passion of understanding and creating intelligent systems, and recently contributed to creating AlphaGo, the first computer program to defeat a human professional player in the full-sized game of Go.
My research interests revolve around multi-agent learning. Over the past 20 years machine learning has evolved to become the leading paradigm in artificial intelligence. In my work, I explore how we can build more and more intelligent systems or agents that learn from experience.
Multi-Agent learning systems are of fundamental importance for this endeavour four three reasons:
- Much of the world’s complexity can be understood in terms of multiple agents interacting - ranging from cells in our bodies to interactions among people to interactions of organisations as large as nation states. Succeeding in such a multi-agent world requires specific skills. Most importantly, artificial agents need to be able to cooperate with other agents - including humans, other artificial agents, and entire organisations. How can we train artificial agents to acquire these social skills?
- Many entities that we think of as intelligent agents are composed of subagents. This includes ourselves as being composed of many different cells, but also teams or organisations that are composed of many different people, yet act as as coherent agents. Even the global market economy can be viewed as an agent composed of many firms that interact through supply chains. How can we harness the power of multi-agent architectures for artificial intelligence?
- Human intelligence - arguably the highest level of intelligence observed in Nature - did not evolve in isolation but as part of a rich evolving eco-system in an interplay of cooperation and competition at the level of species, tribes and individuals. If we wish to design learning agents that acquire their abilities through interaction and by learning from experience, we not only need to design powerful learning algorithms, but also create rich multi-agent worlds in which those learning agents can interact. What will multi-agent training systems look like that can create automatic learning curricula to foster ever greater intelligence in artificial agents?
My recent AAMAS 2020 keynote Automatic Curricula in Deep Multi-Agent Reinforcement Learning gives a great introduction to our team’s work on multi-agent learning systems.
For a shorter introduction, you can also check out my TEDx talk The human pursuit of artificial intelligence.